Efficient Joined Pyramid Network Applied to Semantic Segmentation for GPU Embedded System

Abstract

Fully Convolutional Networks (FCN) are the best methods for semantic segmentation. However, these networks are implemented in computers with high processing capabilities since they are computationally complex. For this reason, many strategies have emerged in the literature to propose novel efficient FCN for embedded applications. In order to contribute with this effort, we propose a network called Efficient Joined Pyramid Network (EJPNet), which is an efficient FCN that reduces the number of activations maps with pointwise convolutions in the encoder and also reduces the number of parameters with an Efficient Joint Pyramid Upsampling (JPU) decoder. EJPNet and other four state of the art efficient semantic segmentation methods were implemented on a Graphic Processing Unit (GPU) embedded system to obtain the performance in precision, number of parameters, memory footprint, Floating Point Operations per Second (FLOPS), and time processing. According to the results, EJPNet not only has the best number of parameters, FLOPS, and time processing, but also this network has one of the best precisions. Then, considering these results, EJPNet is a feasible semantic segmentation method for embedded systems applications.